FlagEmbedding
FlagEmbedding copied to clipboard
FlagOpen
when finetune visualized-bge-m3,expected scalar type Half but found Float
run code:
CUDA_VISIBLE_DEVICES=0 nohup torchrun --nproc_per_node 1 --nnodes 1 --node_rank 0 --master_addr "localhost" --master_port 10003 \
run_ds_cirr.py \
--output_dir ./output/ \
--bge_model_name_or_path ./BAAI/bge-m3 \
--visual_model_name_or_path EVA02-CLIP-L-14 \
--dataloader_num_workers 1 \
--train_data ./ \
--train_data_image ./cirr \
--train_group_size 5 \
--learning_rate 2e-7 \
--fp16 \
--per_device_train_batch_size 4 \
--dataloader_drop_last True \
--normlized True \
--temperature 0.02 \
--logging_steps 10 \
--num_train_epochs 50 \
--negatives_cross_device \
--train_text_tower True \
--train_vision_tower True \
--resume_path ./BAAI/visualized-bge-m3/Visualized_m3.pth \
--save_steps 1000 \
--deepspeed ./deepspeed/ds_config1.json \
--gradient_checkpointing \
> ./train.log &
error message:
[rank0]: Traceback (most recent call last):
[rank0]: File "/home/kmap/zhangquan/bge-visualized/FlagEmbedding/research/visual_bge/visual_bge/downstream/run_ds_cirr.py", line 170, in <module>
[rank0]: main()
[rank0]: File "/home/kmap/zhangquan/bge-visualized/FlagEmbedding/research/visual_bge/visual_bge/downstream/run_ds_cirr.py", line 161, in main
[rank0]: trainer.train()
[rank0]: File "/home/kmap/miniconda3/envs/bge-v/lib/python3.10/site-packages/transformers/trainer.py", line 1938, in train
[rank0]: return inner_training_loop(
[rank0]: File "/home/kmap/miniconda3/envs/bge-v/lib/python3.10/site-packages/transformers/trainer.py", line 2279, in _inner_training_loop
[rank0]: tr_loss_step = self.training_step(model, inputs)
[rank0]: File "/home/kmap/miniconda3/envs/bge-v/lib/python3.10/site-packages/transformers/trainer.py", line 3318, in training_step
[rank0]: loss = self.compute_loss(model, inputs)
[rank0]: File "/home/kmap/zhangquan/bge-visualized/FlagEmbedding/research/visual_bge/visual_bge/downstream/trainer.py", line 47, in compute_loss
[rank0]: outputs = model(**inputs)
[rank0]: File "/home/kmap/miniconda3/envs/bge-v/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1739, in _wrapped_call_impl
[rank0]: return self._call_impl(*args, **kwargs)
[rank0]: File "/home/kmap/miniconda3/envs/bge-v/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1750, in _call_impl
[rank0]: return forward_call(*args, **kwargs)
[rank0]: File "/home/kmap/miniconda3/envs/bge-v/lib/python3.10/site-packages/deepspeed/utils/nvtx.py", line 20, in wrapped_fn
[rank0]: ret_val = func(*args, **kwargs)
[rank0]: File "/home/kmap/miniconda3/envs/bge-v/lib/python3.10/site-packages/deepspeed/runtime/engine.py", line 2063, in forward
[rank0]: loss = self.module(*inputs, **kwargs)
[rank0]: File "/home/kmap/miniconda3/envs/bge-v/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1739, in _wrapped_call_impl
[rank0]: return self._call_impl(*args, **kwargs)
[rank0]: File "/home/kmap/miniconda3/envs/bge-v/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1845, in _call_impl
[rank0]: return inner()
[rank0]: File "/home/kmap/miniconda3/envs/bge-v/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1793, in inner
[rank0]: result = forward_call(*args, **kwargs)
[rank0]: File "/home/kmap/zhangquan/envs/FlagEmbedding/research/visual_bge/visual_bge/modeling.py", line 340, in forward
[rank0]: query_reps = self.encode_mm(mm_it_query[0], mm_it_query[1])
[rank0]: File "/home/kmap/zhangquan/envs/FlagEmbedding/research/visual_bge/visual_bge/modeling.py", line 274, in encode_mm
[rank0]: encoder_outputs = self.bge_encoder(
[rank0]: File "/home/kmap/miniconda3/envs/bge-v/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1739, in _wrapped_call_impl
[rank0]: return self._call_impl(*args, **kwargs)
[rank0]: File "/home/kmap/miniconda3/envs/bge-v/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1750, in _call_impl
[rank0]: return forward_call(*args, **kwargs)
[rank0]: File "/home/kmap/miniconda3/envs/bge-v/lib/python3.10/site-packages/transformers/models/xlm_roberta/modeling_xlm_roberta.py", line 522, in forward
[rank0]: layer_outputs = layer_module(
[rank0]: File "/home/kmap/miniconda3/envs/bge-v/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1739, in _wrapped_call_impl
[rank0]: return self._call_impl(*args, **kwargs)
[rank0]: File "/home/kmap/miniconda3/envs/bge-v/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1750, in _call_impl
[rank0]: return forward_call(*args, **kwargs)
[rank0]: File "/home/kmap/miniconda3/envs/bge-v/lib/python3.10/site-packages/transformers/models/xlm_roberta/modeling_xlm_roberta.py", line 411, in forward
[rank0]: self_attention_outputs = self.attention(
[rank0]: File "/home/kmap/miniconda3/envs/bge-v/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1739, in _wrapped_call_impl
[rank0]: return self._call_impl(*args, **kwargs)
[rank0]: File "/home/kmap/miniconda3/envs/bge-v/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1750, in _call_impl
[rank0]: return forward_call(*args, **kwargs)
[rank0]: File "/home/kmap/miniconda3/envs/bge-v/lib/python3.10/site-packages/transformers/models/xlm_roberta/modeling_xlm_roberta.py", line 338, in forward
[rank0]: self_outputs = self.self(
[rank0]: File "/home/kmap/miniconda3/envs/bge-v/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1739, in _wrapped_call_impl
[rank0]: return self._call_impl(*args, **kwargs)
[rank0]: File "/home/kmap/miniconda3/envs/bge-v/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1750, in _call_impl
[rank0]: return forward_call(*args, **kwargs)
[rank0]: File "/home/kmap/miniconda3/envs/bge-v/lib/python3.10/site-packages/transformers/models/xlm_roberta/modeling_xlm_roberta.py", line 267, in forward
[rank0]: context_layer = torch.matmul(attention_probs, value_layer)
[rank0]: RuntimeError: expected scalar type Half but found Float
deepspeed config:
{
"zero_optimization": {
"stage": 1,
"reduce_bucket_size": 5e8
},
"fp16": {
"enabled": "auto",
"loss_scale": 0,
"initial_scale_power": 10,
"loss_scale_window": 1000,
"hysteresis": 2,
"min_loss_scale": 1
},
"bf16": {
"enabled": "auto",
"loss_scale": 0,
"initial_scale_power": 10,
"loss_scale_window": 1000,
"hysteresis": 2,
"min_loss_scale": 1
},
"optimizer": {
"type": "AdamW",
"params": {
"lr": "auto",
"betas": "auto",
"eps": "auto",
"weight_decay": "auto",
"torch_adam": true
}
},
"scheduler": {
"type": "WarmupDecayLR",
"params": {
"warmup_min_lr": "auto",
"warmup_max_lr": "auto",
"warmup_num_steps": "auto",
"total_num_steps": "auto"
}
},
"gradient_accumulation_steps": "auto",
"gradient_clipping": "auto",
"steps_per_print": 1000,
"train_batch_size": "auto",
"train_micro_batch_size_per_gpu": "auto",
"wall_clock_breakdown": false
}
when I remove:
--deepspeed ./deepspeed/ds_config1.json
--gradient_checkpointing
it can run,but got big grad_norm:
0%| | 170/70560 [03:00<20:14:59, 1.04s/it]
{'loss': 1.784, 'grad_norm': 750.8978881835938, 'learning_rate': 1.9955215419501133e-07, 'epoch': 0.02}
0%| | 170/70560 [03:00<20:14:59, 1.04s/it]06/23/2025 16:47:18 - INFO - root - task types: mm2mm; loss: tensor(0.8470, device='cuda:0', grad_fn=<NllLossBackward0>)
0%| | 180/70560 [03:10<19:49:13, 1.01s/it]
{'loss': 1.5542, 'grad_norm': 1556.9332275390625, 'learning_rate': 1.9952380952380952e-07, 'epoch': 0.03}
0%| | 180/70560 [03:10<19:49:13, 1.01s/it]06/23/2025 16:47:28 - INFO - root - task types: mm2mm; loss: tensor(0.5215, device='cuda:0', grad_fn=<NllLossBackward0>)
0%| | 190/70560 [03:20<19:54:49, 1.02s/it]
{'loss': 1.4019, 'grad_norm': 2484.17724609375, 'learning_rate': 1.994954648526077e-07, 'epoch': 0.03}
0%| | 200/70560 [03:30<19:54:33, 1.02s/it]
{'loss': 1.9826, 'grad_norm': 1053.1171875, 'learning_rate': 1.9946712018140588e-07, 'epoch': 0.03}
0%| | 209/70560 [03:39<19:30:36, 1.00it/s]06/23/2025 16:47:58 - INFO - root - task types: mm2mm; loss: tensor(1.5333, device='cuda:0', grad_fn=<NllLossBackward0>)
0%| | 210/70560 [03:40<19:49:50, 1.01s/it]
{'loss': 1.5291, 'grad_norm': 3596.712890625, 'learning_rate': 1.9944160997732425e-07, 'epoch': 0.03}
0%| | 219/70560 [03:50<19:55:06, 1.02s/it]06/23/2025 16:48:08 - INFO - root - task types: mm2mm; loss: tensor(1.6844, device='cuda:0', grad_fn=<NllLossBackward0>)
0%| | 220/70560 [03:51<20:04:43, 1.03s/it]
{'loss': 1.614, 'grad_norm': 7319.1962890625, 'learning_rate': 1.9941326530612244e-07, 'epoch': 0.03}
is it correct?
change deepspeed config and model.dtype from torch.float32 to half,it works,but still get big grad_norm
"fp16": {
"enabled":true,
"loss_scale": 0,
"initial_scale_power": 10,
"loss_scale_window": 1000,
"hysteresis": 2,
"min_loss_scale": 1
},
"bf16": {
"enabled": flase,
"loss_scale": 0,
"initial_scale_power": 10,
"loss_scale_window": 1000,
"hysteresis": 2,
"min_loss_scale": 1
},
1%|▏ | 920/70560 [13:54<17:13:18, 1.12it/s]
{'loss': 1.8876, 'grad_norm': 8041.6025390625, 'learning_rate': 9.871453272485047e-08, 'epoch': 0.13}
1%|▏ | 930/70560 [14:03<17:16:53, 1.12it/s]
{'loss': 1.8672, 'grad_norm': 8668.6171875, 'learning_rate': 9.870035998752798e-08, 'epoch': 0.13}
1%|▏ | 940/70560 [14:12<17:15:02, 1.12it/s]
{'loss': 2.1783, 'grad_norm': 2928.843017578125, 'learning_rate': 9.86861872502055e-08, 'epoch': 0.13}
1%|▏ | 950/70560 [14:21<17:11:35, 1.12it/s]
{'loss': 1.7931, 'grad_norm': 1147.198974609375, 'learning_rate': 9.867201451288301e-08, 'epoch': 0.13}